| Recently,online learning has become an important way for learners to learn knowledge.It not only breaks the constraints of learning time and place but also helps learns more.Online learning environments have attracted millions of learners to learn and expand their academic knowledge and skills at an unprecedented scale,especially in MOOCs environments.In these platforms,learners’own learning styles and study behaviors lead to generating a high volume of heterogeneous and multi-variate data about learners’profiles,learning outcome information,and records of their interactions with various course resources,which are a stored form of web logs.Interestingly enough,in online learning environments,learners’data are avail-able at varying levels（i.e.,fine-grain to coarse-grain）and can be structured in dif-ferent formats（e.g.aggregated and temporal）.All these data could be employed not only to detect problems from the observed data,but also to predict learners’behaviors and learning outcomes.Plus,the diversity of learners and the complex-ity of learning behaviors under various learning styles,new,unique problems and challenges are posed to machine learning and deep learning models to enable it to explore what lies behind the data.By using big data analysis,machine learning models and visualization technol-ogy,it becomes possible to collect learners’learning data from learning platforms,then process and model them,explore various learning patterns and behaviors be-hind this data,and then improve learning process based on making suitable de-cisions from these technologies.Additionally,it also provides a personalized help for learners in different learning situations,and improves the efficiency of education management and learning.Therefore,learners’learning behavior analysis has become one of the hot issues in current educational application research.In this context,many relevant research papers have been published for educa-tional data analysis based on machine learning methods.These studies proposed analytical models that can analyze learners’study behaviors,predict their perfor-mance,and improve education management and learning efficiency.At the same time,the analysis and mining of educational data have also promoted the optimiza-tion of machine learning and deep learning models.Nevertheless,these models face some limitations and there are still the following problems:（1）Existing early prediction models which focus mostly on the prediction of dropping out risk,but it is obviously of further research significance to evaluate quantitatively the dropout risk of learners,which will provide an objective basis for more personalized interventions and supports for those who need help the most.（2）Most prediction models are constructed based on feature engineering,which means that,specific features that can lead to different recognition results are taken as prediction features.Thus,these models cannot apply to data with different ar-chitectures from different platforms or to new datasets.Additionally,such models neglected the time dimension that could better reveal the effect of learners’perfor-mance during their studying.（3）The online learning data is heterogeneous.Labeling the level of learners’en-gagement based on their learning behavior is laborious,costly and time-consuming,especially the classification which employ nominal class labels.On the other hand,supervised algorithms rely on a large amount of labeled data that can be used to achieve more accuracy at predicting outcomes.（4）The results of the prediction models are represented by numerical values,which means that they lack a visual representation to visualize how the learner works during the learning process and what the learner does while watching the video（e.g.,instructors and video providers may want to know how learners engage the lecture videos,and which pieces in each video attract learners and which pieces are skipped）.In response to the above problems,the present research studies the optimiza-tion of the machine learning models by proposing accurate and efficient prediction models to improve the dropout rate and the accuracy of the prediction of learning re-sults.Additionally,it enhances the interpretability of the results of the models with visualization methods.In contrast,it studies the impact of machine/deep learning methods on the learning analytics field in order to understand the process of learn-ing in online learning environments.Additionally,a learning behavior visualization system of learners is designed to explore how they learn differently,plus identifying their behavioral patterns and classifying learners on the basis of their engagement.The main work and contributions of this dissertation are as follows:（1）Based on the logistic regression model and the input-output hidden Markov method,we proposed an early dropout prediction model,termed EIARS,which aimed to optimize dropout prediction performance by identifying the earliest possi-ble time frame throughout the consecutive weeks of the course;and to provide an approach to personalize and prioritize intervention for more at-risk learners through using individual dropout probabilities.In this model,an algorithm to extract im-portant features related to the dropout based on features engineering strategy was proposed.Thus,depending on logistic regression and L2-regularization,we proposed a transfer learning algorithm（LR-SM）,which learns across weeks simultaneously to balance smoothness and accuracy by adding a new regularization term that takes into account the probabilities in sequential weeks and then minimizes their differ-ence.The experimental results show that the proposed model has a higher accuracy rate than other traditional machine learning models in addition to visualizing results of the proposed model in order to explore and identify the learners with the highest dropout risk.（2）As the fixed-term dropout prediction suggests,current models depend on feature engineering strategy（manual extraction）.This process is costly,time-consuming,and unextendible to new datasets from different platforms or different courses with different characteristics.Therefore,we proposed a novel Deep On-line Prediction（DOP）model that can automatically extract features from raw data stored in online learnings platforms.This DOP model takes into consideration the sequential relationship of learners’behavior towards dropout and class imbalance problem in dataset.The model structure is built upon a data transformation al-gorithm（DT）,model combined of CNN and LSTM with a custom loss function（CLF）to predict whether each learner will drop out or complete the course.Data transformation algorithm（DT）converts raw logs into meaningful data according to time of learners’data records and behavior dimensions to a two-dimensional ma-trix fed as inputs to CNN layers.CNN layers extract significant features from each matrix automatically to obtain the micro-level characteristics that are effective for dropout prediction.Due to a learner’s behavior at a given period,which is influ-enced by his historical behavior,we took into account time-series properties between different time slice features through LSTM.Moreover,to solve the problem of er-ror classification on prediction,we proposed a custom loss function（CLF）relying on a cost-sensitive technique by statistically calculating costs inferred according to learner’s activities and the actual days spent in the course.Therefore,it tunes differ-ent weights for different classes,solves the impact of unbalanced data on prediction results,and improves the performance of dropout prediction.（3）Many different machine learning algorithms have been employed to classify in online education data.These algorithms require enormous amounts of labeling data for classification,which is unrealistic for online learning data.Since online learning data is heterogeneous,the main challenges associated with this data are lacking labeled and classification using nominal class labels,rather than ordinal ones.Additionally,the current classification approaches fail to accurately classify learners into specific behavioral patterns based on their own learning styles.Mean-while,such approaches are not able to analyze high-dimensional learning behaviors among learners according to their interaction with course content.Therefore,in the third work,we proposed a model based on Residual Graph Con-volutional Network（Res-GCN）,as a semi-supervised classification task to classify and identify learners’engagement levels to various behavioral patterns.Firstly,we built a heterogeneous knowledge graph（HKG）representing learners,course videos,and knowledge concepts as entities.The model works to solve the gradient van-ishing problem in deep GCNs by adding a residual connection between different layers of the deep GCNs to ensure that the L+1 layer network contains more image information than the L layer.Thus,the Res-GCN model intrinsically works for the heterogeneous knowledge graphs as a semi-supervised node classification task to classify learners’engagement levels.In this study,we hypothesized four behav-ioral patterns while learners learned from the course,namely（“High-engagement”,“Normal-engagement”,“At-risk”,and“Potential-At-risk”）.We evaluated the effectiveness of our approach on a real-world dataset,the dataset lacked nominal class labels as a ground truth.Therefore,we proposed a label function（LF）based on the data programming method to label datasets instead of manual labeling,in which both input and output data are labeled for classification to provide a learning foundation for future data processing.Results of the experiment showed that our approach can achieve a much higher accuracy of learner classification than what the traditional machine learning approaches and related works can achieve.（4）In order to improve the interpretability and transparency of the machine learning model,a visual analysis system（MOOC-ASV）was designed,which is aimed at visualizing learners’interaction with course videos and how these interactions hap-pen.Besides,it helps teachers to understand learners’performance,and improving teaching strategies,and help learners to improve their learning efficiency.In this system,based on graph theory,we first proposed a learner-video bipartite graph Glvbased on learners’engagement with course videos.Thus,to cluster learners ac-cording to their’engagement with course videos on a graph,the Louvain method to detect communities was employed on graph Glv,using the edge weights and default resolution of 1.0.Second.The struc2vec was used to discover structural identities for learners’node by identifying direct connections the node and the number of its direct connections had.Then,the k-means technique was applied to the created vectors by struc2vec.This enabled us to identify a structural classification for each node.Nodes that have a similar structural identity cluster together in one cluster.In addition,a novel glyph design to classify student performance using the utilization rate（UR）was proposed as an indicator to an evaluation of their performance.In our designed system integrated well-established visualization techniques and several novel visual designs that can highlight issues related to course content. |
